KR20210079264A - Methods and systems for defining emotional machines - Google Patents

Abstract

Disclosed is a method of training an intelligent agent comprising generating a personality matrix, combining the cognitive bias matrix with the personality matrix, and generating a behavioral function for a situation based on the combined cognitive bias matrix and the personality matrix. do.

Description

Methods and systems for defining emotional machines

Aspects of the present disclosure relate to expert systems, and more particularly, to the development of expert systems and machine learning that use psychological and sociological information for greater behavioral replication.

Intelligent systems, which mean machines and network services work together, could have a greater ability to memorize and compare auditory, visual and other sensory cues than humans.

Starting with sight, camera technology suggests that intelligent systems (IS) can do much more than humans can do - behind, up, down, far away, in little light, invisible to the human eye, but detected by some animals. It has been improved to a level where everything can be seen in the same frequency range as infrared and ultraviolet that can be. In fact, the IS can see other electromagnetic waves such as X-rays, microwaves, and radio waves. To understand what humans see, IS will know the limits of human vision and will consider what they would see if they were humans, but could use other data if desired to create a superhuman vision (and superhuman memory and accuracy of what it saw). do.

As for hearing, high-resolution binaural recordings are already very similar to what humans hear, but why stop there? Systems are being developed that can separate individual sounds or stems from performance so that the IS can observe individual elements of the auditory environment.

How does it taste and smell? "Since 1982, research has been underway to develop technologies, commonly referred to as electronic noses, that can detect and recognize odors and flavors." Since then, work has progressed significantly as MIT's Andreas Mersin and Shu Kwang-Jang, who won the DARPA award for their Nano-Nose, better understand how olfactory perception works.

The ability to identify objects by touch is one of the key functions of human sensing. BioTac tactile sensors have “the ability to discriminate and identify objects based on their compliance, texture and thermal properties, not only with performance, but sometimes even better than human perception.”

Irun Wang of Stanford School, Michal Kozinski, conducted a 2017 study that demonstrated the ability of deep neural networks to detect sexual orientation from facial images with greater accuracy than human controls.

Some systems can now read emotional and behavioral cues more accurately than humans. Scientists at the Massachusetts Institute of Technology's Affective Computing Lab have discovered that the Emotion Facial Action Coding System, developed in the 1980s by Paul Ekman and Wallace V Friesen, but is much more advanced; EMFACS) evolved from "Artificial Emotional Intelligence" to market.

It is in this context that aspects of the present disclosure arise.

The teachings of the present invention may be readily understood by consideration of the following detailed description in conjunction with the accompanying drawings:
1 is a schematic overview of components of an IS in accordance with aspects of the present disclosure;
2 is a diagrammatic diagram of human behavior as a series of layers in accordance with aspects of the present disclosure.
3 is a block diagram illustrating a Meyers Briggs Personality Typing continuum in accordance with aspects of the present disclosure.
4 is a diagram view of five personality traits in accordance with aspects of the present disclosure.
5 is a block diagram illustrating contextual reference and sensing inputs and outputs in accordance with aspects of the present disclosure.
6 is a schematic diagram of parameters of a power generation filter in accordance with aspects of the present disclosure.
7 is an exemplary diagram of elements of a relational filter in accordance with aspects of the present disclosure.
8 is an exemplary diagram of elements of behavior masks in accordance with aspects of the present disclosure.
9 is an exemplary diagram of a mental stack including behavior-specific functions in accordance with aspects of the present disclosure.
10 is a diagram illustrating detailed elements of behavior-specific functions in accordance with aspects of the present disclosure.
11 is a block diagram showing a complete stack from DNA via behavior based observation functions in accordance with aspects of the present disclosure.
12 is a diagram illustrating a connection between a reference persona and individual instances of an intelligent system in accordance with aspects of the present disclosure.
13 is a block diagram illustrating mapping of behavioral biases and various filters and masks in accordance with aspects of the present disclosure.
14 is a diagram showing how an expert system is used to map behavioral or cognitive biases in accordance with aspects of the present disclosure.
15 is a block diagram illustrating the layers constituting a personality criterion in accordance with aspects of the present disclosure.
16 is a table showing MBTI weights in accordance with aspects of the present disclosure.
17 is a block diagram illustrating cultural hierarchies in accordance with aspects of the present disclosure.
18 is a block diagram illustrating behavior collection in accordance with aspects of the present disclosure.
19 is a block diagram illustrating how a contextual criterion is mapped to behavioral biases to generate a contextual bias set in accordance with aspects of the present disclosure.
20 is a block diagram showing a mapping of psychological parameters to behavioral biases that weight each in accordance with aspects of the present disclosure.
21 is a table illustrating an example of a behavioral bias matrix constituting an IS instance for a situation in accordance with aspects of the present disclosure.
22 is a block diagram illustrating input biases for contextual environments for each IS in accordance with aspects of the present disclosure.
23 is a table illustrating an example of a personality matrix in accordance with aspects of the present disclosure.
24 is a table showing an alternative view of reference personality matrices in accordance with aspects of the present disclosure.
25 is a block diagram illustrating capturing and analyzing behavioral data in accordance with aspects of the present disclosure.
26 is a block diagram showing a view of a social taxonomy in accordance with aspects of the present disclosure.
27 is a table view of parameters including an overall personality matrix in accordance with aspects of the present disclosure.
28 is an unclassified matrix view of a matrix used to describe personalities in accordance with aspects of the present disclosure.
29A is a simplified node diagram of a recurrent neural network for use in an intelligent system in accordance with aspects of the present disclosure.
29B is a simplified node diagram of an unfolded recurrent neural network for using an intelligent system in accordance with aspects of the present disclosure.
29C is a simplified diagram of a convolutional neural network for use in an intelligent system in accordance with aspects of the present disclosure.
29D is a block diagram of a method for training a neural network in the development of an intelligent system for aspects of the present disclosure.
30 is a block diagram illustrating training of a generative opposing pair neural network in an intelligent system in accordance with aspects of the present disclosure.
31 illustrates a block diagram of an intelligent agent system in accordance with aspects of the present disclosure.

Although the following detailed description contains many specific details for purposes of illustration, those skilled in the art will understand that many modifications and variations to the following details are within the scope of the present invention. Accordingly, the exemplary embodiments of the invention described below are set forth without loss of generality and without imposing limitations on the claimed invention.

Introduction

Machines have not yet mastered the art of human interactions. Although many chat bots have tricked users on a regular basis, communication with machines is often repetitive, logical and clearly inhuman. Humans generally do not act rationally. We have dozens of cognitive biases. However, these behaviors are "unpredictably irrational". Nothing can stop intelligent machines from behaving "irrationally" in the same ways humans do.

In the field of behavioral economics, many studies have been conducted in the fields of boundary rationality, prospect theory, inter-time selection, nudge theory, behavioral finance and behavioral game theory. As these theories evolve and become more positive, nothing can stop intelligent machines from behaving in the same way humans make economic decisions. Alternatively, nothing can prevent IS-enhanced humans from becoming less irrational.

Aspects of the present disclosure may be subdivided into a number of components that, taken together, will provide a complete system for defining and building intelligent systems or ISs. This IS includes, but is not necessarily limited to, devices, networks, storage, data structures, processing, algorithms, inputs, outputs, and deep neural networks, convolutional neural networks, recombinant neural networks, expert system generative alleles. It includes some or all of a variety of artificial intelligence technologies, including networks and artificial neural networks using training and/or inference. The goal is to teach ISs - including instances of fully humanoid robots in simple chat bots to make them behave more like humans. To this end, we will evaluate human responses, psychological, sociological, physical - and the ways in which an intelligent system (IS) can imitate them.

The problem can be broken down into several components. Referring to FIG. 1 , an IS 100 can record 101 the same inputs that a human has with sight 102 , hearing 103 , touch 104 , taste 105 and smell 106 . it can be seen that there is The IS can then analyze and calculate 108 these inputs—in near real time—and perform a response 109 using haptics, speech generation, and robotics. Responses constructed by ISs should be able to mimic human responses in ways that are indistinguishable from humans by other humans, and potentially much more empathic (or Machiavellian). Much of this disclosure will address the social and psychological aspects of these understandings and responses. Finally, how can these systems be used to create virtual environments and react in the real world? Let's take a closer look at the high-level architecture.

inputs

Looking more closely at the inputs, intelligent systems (meaning machines and network services working together) may have greater ability to memorize and compare auditory, visual and other sensory cues than humans.

Time

Starting with vision 102 , certainly camera technology has advanced to the point where the IS can see everything a human can do and much more in some aspects of the present disclosure. Back, above, below, far away, in the same frequency range as infrared and ultraviolet, which has little light and is invisible to the human eye but can be detected by some animals. In fact, the IS can see (through the use of these sensors) other electromagnetic waves such as X-rays, microwaves and radio waves. To understand what humans see, IS will know the limits of human vision and will consider what they would see if they were humans, but could use other data if desired to create a superhuman vision (and superhuman memory and accuracy of what it saw). will be programmed to Using machine vision and/or object recognition, IS can detect and classify objects in the physical world, including humans, to formulate realistic human responses.

ear

With regard to hearing 103, high-resolution binaural recordings very closely mimic what and how humans already hear. According to aspects of the present disclosure, IS can record sound into objects, just as systems such as Dolby Atmos reproduce audio as individual objects, capture sound in a stereo or surround field and convert it into individual elements. Detachable General Harmonics, Celemony, Red Fill and other systems from companies such as Sony are being developed. For example, without limitation, a person receiving IS or IS assistance may want to listen to a symphony and only hear the first violin or French horn, and only these instruments can be isolated, so that the IS essentially uses fine-grained control (analysis of individual components) (using spectral decomposition methods such as ICA) can turn the real world into a recording studio. In some embodiments, the IS may be integrated into human biology, and the human may simply think, but at a concert "I wish the French horn got a little louder" and you (or your cyborg construct) elements) can "change the mix". With this integration, not only can a person remember everything, but you can listen to it again in a different mix.

taste and smell

The electronic nose can recognize odors and tastes perceived by humans, as well as those recognized by other animals, such as dogs. Different people react differently to different smells and tastes, but knowing how I smell and how you smell (used as a verb) is a pure machine learning data analysis problem. Since this is simply a matter of identification, quality is not involved. ISIS can certainly learn to recognize the tastes and smells of foods preferred by other cultures and by different individuals. In fact, when the IS eats food (chews and swallows) with the right hardware, it can also learn textures (or store them in memory). You won't necessarily have to digest food for energy (though it may be designed). However, a machine can be trained to recognize (or be perceived by outsiders to recognize) all the subtleties of appreciation for food and drink as a human or any particular human being. A Jonathan Gold machine with an IS instance could be hired to help judge the menu in a restaurant. Alternatively, restaurant chains can dynamically adjust seasonings based on customers' tastes through IS chefs with the appropriate training and hardware.

sense of touch

The ability to identify objects by touch is one of the key functions of human sensing. Touch sensors can be used to detect initial contact, contact position, slip, radius of curvature, and edges, as well as determine triaxial forces that allow skillful handling of objects with unknown properties at unknown locations. As with the other senses, for the sake of this disclosure, we will assume that high-resolution touch—both giving and receiving—will become more available over time.

Analysis by Behavior

The heart of this initiation has to do with the psychological and emotional aspects of our being. Humans are quite good at recognizing the emotional states and dispositions of other humans. A person can distinguish whether someone is angry, happy, or dozens of other emotions. Analysis of other people's emotions and tendencies is not only about what people say, but also about body language and facial expressions - including subtle expressions, voice tone and pitch, smell and physical observations such as flushed skin, goosebumps, tears, etc. based on reading. People's interpretations of this are somewhat colored according to their experiences, preconceived notions, and expectations. Paranoid people think that other people deliberately do bad things to them, but there are more subtle versions that map one person's expectations to another's behavior and the environment in general. This disclosure will look at how IS can or better understand not only humans, but also emotions and behavior-specific tendencies.

According to aspects of the present disclosure, the IS may take into account the emotions of a person interacting with the IS, such as sadness, joy, anger, empathy, or frivolity. Moreover, an IS in accordance with aspects of the present disclosure may interact with and continuously learn from other inputs such as televisions, billboards, sales in showroom windows, price movements or crowd behavior.

Before reacting to emotional inputs, the IS must be able to read and "understand" them. This disclosure does not focus on primitives that read emotions. Instead, it will be assumed that aspects of the present disclosure collect all elements and analyze what they mean for the psychology and sociology of the environment, although other techniques may be used to capture the underlying primitives.

reactions

According to aspects of the present disclosure, the IS may exhibit an appropriate response to a given emotional input. Depending on the input and the circumstances surrounding the input, the IS may respond with emotions such as empathy, anger, contempt, and groupthink, for example. In some implementations, the IS may respond to input and situations with a decision to act, eg, to buy, sell, or to act, eg, to clean or cook. When intelligent machines are trained to read emotions with physical micro cues (facial expressions and other body language, smell touches, etc.), intelligent devices (e.g. robots) can process physical actions such as facial expressions, smells, sweat, etc. They can be trained to imitate, allowing humans to experience empathy, contempt, and so on.

When eaten with an IS dating status, she/he can recognize the taste and smell of food and react appropriately - even taking into account the person's known preferences. At the end of a heavy meal, she/he may feel full, lethargic, and become intoxicated and sexually aroused.

Humans generally do not act rationally. We have dozens of cognitive biases. However, these behaviors are "unpredictably irrational". Nothing can stop intelligent machines from behaving "irrationally" in the same ways humans do.

The steps necessary to make rational, even highly rational, empathic (or less ideally operant) responses are analyzed in detail. As shown in Figure 2, human behavior can be viewed as a series of hierarchies. There are many inputs to the sensory analysis engine 200 . First, there is the sensory recording 201 as described above. There are also environmental factors 202, which will be discussed in more detail later. The other key elements needed to construct the response calculation 203 are the personality elements 204 that generally shape our behavior. Then, the behavioral mask 205 - refers to both the mask we wear when presenting ourselves to others and the mask in computer programming, which refers to filtering the various bits or bytes and removing them from the output. Thereafter, behavior-specific functions 206 - using functions in a mathematical sense that have an algorithm or set of operations that operate on a set of inputs to produce a desired output, but which apply to psychological decision making - will then be discussed.

Situational criteria

Contextual criteria are the basic personality structures that IS can bring to any situation or interaction. Situational criteria have three components: The first is a basic personality type or personality component. In humans, this is mostly a result of genetics and infancy and often represents a fundamental view of humans (eg, children who have been abused, usually not learning to trust). In terms of replicating human responsiveness, the next developmental layer is the developmental filter. The developmental filter is a cultural and social overlay on our basic personality. It is driven by our social and cultural environment, which may include family, community, friends, etc. The third element is the relationship filter. There are filters that act on us based on context. It reflects existing relationships with current places and people.

As used herein, basic personality components refer to the quantification and analysis of basic human characteristics. There is no doubt that these traits have a genetic component, and in the future they will undoubtedly be used in the analysis of basic personality factors. In some embodiments, basic personality components are limited to a psychological (and data) approach to basic personality analysis. In alternative embodiments, even genetic traits and predispositions are considered, and these genetic traits can be applied using genetic guidance and potential for specific personality traits due to genetic code markers.

Any personality modeling system in the art can be used as the basic personality elements. For example, and without limitation, personality types are generally described in the psychological literature using either of two different models. One model is the Meyers Briggs Personality Typing, which is based on the Jungian architypes and subdivides the personality into 16 binary combinations, as shown in Figure 3:

- Extroverted (301) -> Introverted (302)

- Sensing(303) -> Intuitive(304)

- Thinking(305) -> Feeling(306)

- Judgment type (307) -> Recognition type (308)

Another common personality analysis tool is the Five Personality Traits or Five Factors mode, originally based on the work of Ernest Tufts and Raymond Crystal, later developed by JM Dickman and Lewis Goldberg, and by some, the underside of all personality traits. It is believed to represent the basic structure in

As shown in FIG. 4 , the five major personality traits 400 are generally described as follows:

- Openness to Experience (Creative/Curious vs. Consistent/Cautious) (401)

- Integrity (efficient/organized vs. laid-back/inattentive) (402)

- Extraversion (divergent/active versus isolated/introverted) (403)

- Friendly (kind/sympathetic vs. defiant/separate) (404)

- Neurotic (sensitive/nervous vs. safe/confident) (405)

There are numerous other personality continuums, including the following trait studies by personality physiologists. These include, but are not limited to:

- Machiavellianism: An individual who manipulates the behavior of others, often through duality. Machiavellians are often interested in money and power, and make practical use of others in this quest.

- Need for achievement: People who are desperate for achievement want to achieve a lot and set high standards of excellence for themselves. They can work persistently and hard for distant goals.

- Cognitive Needs: People with high cognitive needs find it rewarding to understand things and are willing to put significant cognitive effort into this quest. These individuals enjoy the process of learning and trying to understand new things.

- Authoritarianism: Authoritarians believe in a strict social hierarchy where they are completely obedient to their superiors and expect complete obedience from their subordinates. The authoritarian nature of strictly following rules is very uncomfortable with uncertainty.

- Narcissism: The narcissistic personality is so narcissistic that it produces high levels of vanity, pride, and selfishness. Narcissistic individuals often have trouble empathizing with or appreciating others.

- Self-Esteem: The tendency to evaluate oneself positively. Self-esteem doesn't mean believing that you are better than others, it just means that you are worthy.

- Optimism: The tendency to expect positive outcomes in the future. Optimistic people expect good things to happen and they often get more positive results.

- Emotional empathy: Inability to recognize and classify one's emotions. These people have a hard time recognizing the feelings of others.

Obviously, these personality traits are tendencies. They are not deterministic (or "normative"). Thus, ISTJs in Myers Briggs term are 80% more likely to double-check their shopping carts than ENFPs. At a basic level, these trends will affect the way IS responds in various social situations. In some embodiments, the basic personality of the IS may be established on a continuum along multiple axes using basic personality elements as an axis.

For example, suppose you want to create an IS entity called "Chris?" without limitation. We can choose our gender and sexual preference because it affects our personality, but there are many more. For example, using Myers Briggs as a basic personality approach, we find that Chris is 75% extraverted, 25% introverted; 43% sensory, 57%; Intuitive; 27% thinking type, 73% feeling type, 48% judgment type, 52% perception type. Similar parameters can be used for the five factor approach. Chris may have an openness to experience scale of 38%, a conscientiousness scale of 72%, an extraversion scale of 81%, and a neuroticism scale of 22%. In other embodiments, the personalities we create are consistent with those known to personality economists or others as a result of analysis of Machiavellianism, achievement needs, cognitive needs, authoritarianism, narcissism, self-esteem, optimism, empathy, and AI behavioral analysis. It can have measures along different continuums as well.

All of this becomes the basic personality type 500 , which is part of the contextual criteria 501 as shown in FIG. 5 . This complete contextual criterion 501 is what the IS brings into the present relationship. In addition to the basic personality types, there are two components. First, the environment in which people grow up is influenced by behavior, i.e., the developmental filter 502 . This includes everything that may have affected a person's development since birth - the relationship with the family, the culture in which the person was raised, the political environment, even the weather. People are also affected by the relationship filter 503 , which is their relationship to those people and the things around them. This includes what the person is saying and what they and the person's history are. For example, the impact of a relationship may depend on whether the interaction with the other person takes place in an office, resort, someone's home, etc.

In some implementations, the representation of a personality type may have 16 basic personality components and may have two factors related to each. Factor 1) is the size of each personality element on a scale up to 100% depending on how strong the IS is on one side or the other, e.g. these are 74% introversion, 25% narcissism, 17% judgmental, 49% Machiavellian etc. Factor 2) is the weight of the importance of each of the 16 personality components within a given situation, for example, how important is narcissism or thinking style to the task at hand or openness to experience. However, aspects of the present disclosure are not limited to these implementations.

This contextual criterion is then influenced by the sensory inputs 504 of sight 505 , hearing 506 , smell 507 , taste 508 and touch 509 . These inputs are then read by the emotion reading filter 510 which generated the baseline cognitive response 511 . These responses are then fed to a behavior-specific function algorithm 512 , which then produces a sensing output 513 .

power generation filter

In accordance with aspects of the present disclosure, the IS may be designed to mimic human behavior, allowing a history of the IS to be generated. When screenwriters write a script, they usually have a "bible" that describes what made the character. A script may not mention where a character was born, but knowing whether a character was raised on a farm in Iowa or in a townhouse in Manhattan can have a huge impact on the way that character behaves and, accordingly, the way an actor will play that character. crazy In the same way, the evolution filter is a character bible for IS. For example, a happy marriage affects a person's behavior very differently than being in an unhappy marriage and being in a happy second marriage after an unhappy first marriage.

As shown in FIG. 6 , factors affecting the power generation filter 600 may be subdivided into several key buckets. Early development 602 includes such things as family factors (size, siblings, parents, etc.), education, financial situation, health, and the like. Here is the data for the current relationship history 603 , and we also want to track the education data 604 , the work history 605 , and any number of other factors 606 that will be determined to be relevant by an expert in the field. can

According to aspects of the present disclosure, an intelligent system may be trained using various kinds of artificial intelligence (deep learning, convolutional neural networks, biologic antagonistic neural networks, etc.), and may essentially provide the learning of other machines out of the box. Therefore, the depth of understanding will increase exponentially. Interactions between any number of ISs can be compared and fine-tuned to actual interactions in human history. Testers can select different characteristics for the IS and run them to see performance differences. It should be noted that not all humans need to be fully accurately represented, and a few documented human interaction histories may suffice.

relationship filter

The following is the relationship between the IS 700 and the person 701 or person 702 and the interacting places 703 , 704 , 705 , as shown in FIG. 7 . There was a lot of psychological analysis in this space. According to aspects of the present disclosure, a corpus of descriptive relationship data may be generated from a psychological survey. The psychological survey will answer questions such as, for example and without limitation: How do you feel when you relate to your boss as opposed to your subordinates (this will be influenced by your basic personality type - are they a code keeper?) Or are you an egalitarian)? What happens to other family members? How do people feel about genetics (for example, about relatives who have not grown up together but have just met)? What is the environment like? Are you most comfortable in an office environment, or more at a bar or someone's home? If the person you're talking to is short or fat, do you feel superior or show off? Do they make eye contact? How sensitive are they to physical cues? How is the environmental noise? In some implementations, the survey may include a question asking the survey participants to define a weight and magnitude for how they feel for each question. In other embodiments, the weights and magnitudes are statistically developed from a collection of psychological surveys answered by humans.

Masks by action

For every action-specific response, there are an infinite number of responses. In the same situation, different people will react differently. Some of the possible affected areas are shown in FIG. 8 . IS (801) includes such things as business associations, romantic attraction, love of intellectual interests or hobbies, where they grew up, what the family environment they grew up in, what their health history was, relationship history, and their psychological types and tendencies. with a person 802 .

Any specific IS should be designed to mask specific responses. ISIS should be designed to mask violence except when protecting others or in self-defense, but still does no harm. These relational filters or behavioral masks are generalized trends based on life history and most important based on background and genetics. Some backgrounds may be recent, some may be older, and more fundamental to the surface.

Action-specific functions

This is where the preceding layers are performed, as shown at a high level in FIG. 9 . Basic personality elements 900 , developmental filter 901 , relational filter 902 , and behavioral mask 903 use these as operators for immediate task response 905 . Tasks are, for example, without limitation; answer questions in conversation, see who has just spoken, decide whether to buy an item or not, choose a different restaurant or time or date if the first choice is not available, offer a substitute to shoppers, or basically anything humans can do today It can be any reaction. One key question is, "How do we want humans to respond?" Humans are not rational actors, and according to aspects of the present disclosure the IS may be configured to mimic irrationality. For example, and without limitation, suppose that someone is angry and there is nothing against it. Perhaps the person has just missed the flight and cannot go to the wedding on time now. If you're a rational actor, you can say that there's nothing you can do about it, all options have been explored and the best thing to do is send an apology text. However, the human actor may say, “Oh my god, what a disappointment! Let's see how we can help. Let's see if there are other ways to get there on time first.” After a proper pause and more empathy, they can say, "Do you wish they would show up the next morning to say goodbye to their honeymoon, or just send a note or text?" The point is that humans care much more about the process than they do about the outcome. Thus, the IS can be designed to mimic human responses by providing appropriate behavioral function layer 904 parameters that provide process awareness.

As another example, suppose a store sets a price for a watch. Humans are subject to a cognitive bias known as "anchoring". As an example, if the price of one item is very high, it makes the other item appear more reasonable even though it is actually still expensive. An IS with appropriate masks or filters in the behavioral function layer 904 can be developed to respond with a human bias as discussed below.

Humans are full of cognitive biases that are part of what makes us human. We believe that humans rationally analyze all factors before making choices or determining values, but the truth is that first perceptions remain in our minds and influence future perceptions and decisions. Another cognitive bias is “apotheosis”. People believe they know when to lie to themselves, but are unaware of the constant nudging we receive from ideas formed in our subconscious. Behavioral psychology is replete with studies demonstrating at least the following named biases:

Priming, hypocrisy, hindsight bias, sharpshooter fallacy, procrastination, bias toward normality, magnetism, availability heuristic, bystander effect, Dunning-Kruger effect, apophenia, brand affinity, argument to authority, argument based on ignorance, scarecrow The fallacy of aggression, the fallacy of personal attacks, the fallacy of an impartial world, the public goods game, the ultimatum game, subjective verification, cult indoctrination, group thinking, emotion heuristic, Dunbar's law, sold-out, self-centered bias, lighting effects, third Self effect, catharsis, adaptation, extinction trigger, supernormal releasers, fundamental attribution error, representativeness heuristic, social roping, learner helplessness, anchoring effect, self-fulfilling prophecy, moment, illusion of control, consistency bias and expectations.

A few simple examples without limitation will suffice to demonstrate how this generally works. Take the Dunning Kruger effect as an example. The Dunning-Kruger effect is a cognitive bias in which people mistakenly estimate their cognitive abilities to be greater than that. This is related to a cognitive bias in fantastic superiority, and comes from people not recognizing their own lack of ability. For ISs designed as learning assistants, adding cognitive biases such as the Dunning-Kruger effect can make the IS more relevant to the user and more enjoyable to learn. For example, suppose a person is learning to program in JavaScript and has a professional programmer who can answer all of your questions. However, this is not fun because of the joy of learning that comes from shared discovery (including programming). An IS with a Dunning Kruger effect bias would share that person's naive passion. This will be true for all tasks as well. By way of example and not limitation, consider learning basketball. It would be no fun to play with a) the person who always made all the baskets, or b) the person who painfully knows his shortcomings. It's more interesting to share a bit of human encouragement and a pure belief in skill development than a machine that can soberly calculate the real odds of making a free throw or slam dunk.

Post-mortem overconfidence bias refers to the tendency of people to perceive events that have already occurred as more predictable than they actually were before they occurred. In this case, an IS with perfect memory and perception would know exactly how often it correctly predicted events, and would say the weather would get worse. Of course, humans make predictions based on feelings like temperature changes, perhaps air pressure ("my arthritis is working"). A person who predicts that it will rain tomorrow and predicts that it will not rain will forget what he predicted, but when it rains, he will remember and say, “I thought it would!”

Figure 10 illustrates the entire stack up to this point. Situation 1000 is created with context-specific criteria (basic personality factors, developmental filters, and relational filters) and behavior-specific masks, which are filtered through behavior-specific functions 1001 such as priming, asymptosis, bias towards normality, etc. and write reaction 1002.

IS may make the following decisions: Cognitive bias can be described, for example, as a function of: Behavior f(Positive Overconfidence Bias) = (Experience programmed on axis #1) (Experience programmed on axis #2 on n) (each the unique event predictability of the axis). For example, in the case of a sailor, the programmed degree of experience is related to what the sailor experiences on the beach's weather axis or what the sailor's weather forecast in the desert. From an IS system perspective, IS is prepared by "situation" - meaning basic personality elements (Myers Briggs, genetic makeup, gender, etc.). By the relationship filter (long-term relationships to people and the environment) modified by the evolution filter (cultural and social upbringing), then further modified by behavioral masks (my current relationship with people involved, social hierarchy, etc.) contextualized). This created the basic context in which functions, such as behavior f (post-hyperconfidence bias), operate.

When discussing intelligent systems within the context of the present disclosure, it is more appropriate to call them cognitive biases, behavioral biases. Training the IS to mimic these behavioral biases will be discussed in a later section.

Before looking at how to train IS using machine learning, it is useful to take a closer look at the entire stack as shown in FIG. 11 . The lowest level input to our behavior is DNA (1100). How DNA influences personality depends on a number of factors created specifically for the IS, or factors related to the person being modeled on the IS. Directly on top of DNA (1100) is how the IS is patterned based on early development (1101). What happens at an early stage of development (abuse, extreme poverty, perfect love, etc.) shape people very deeply and generally permanently and will also affect the character of the IS. There is a gene regulation 1102 that controls behavior over the same period (and continues to a lower level) with a combination of genetic and environmental factors. Above that are the elements that make up the contextual criterion 1103 - consisting first of the basic personality elements 1104, then of the developmental filter 1105 and the relational filter 1106. Further up the stack are behavior masks 1107 that analyze behavioral biases 1108 and impute these biases to the individual 1109 and the environment or behavior and choices 1110 . Action functions 1111 can be used to generate action from all this data. Thereafter, the results of the interaction based on the functions are fed back to the behavior masks as the system continues to learn from that experience.

persona

Each ISI has a path of growth or development. This path is the point at which there are a number of key points but one becomes irreplaceable - that is, when they first interact with humans. For example, and without limitation, suppose you have an IS named Dale who is a personal (concierge) customer care representative. Dale knows the full customer care history of an individual customer across all customer devices from all manufacturers. Dale has a personality developed through contextual criteria. Clients could choose from a number of contextual criteria or choose one for them based on their personality profile. Now, moving forward and based on their interactions with customers and IS, their personalities will evolve. A year from now, Dale will know what customers find interesting about all their purchases and support histories, whether they like chatting or just getting started. This is a persona for individual customers only. If another human customer starts as a rep with the same contextual criteria, that customer's "Dale" IS instance will not remain the same for long. When the IS interacts with this customer, the relationship will develop differently than the relationship between Dale and the first customer. For the purposes of this disclosure, personas developed via contextual criteria (or our virgin personalities) are referred to herein as baseline personas, and each persona “customized by human interaction” is an ISI (Intelligent System Instance). to be. If two customers get married and share a personal customer care agreement with all their devices, that doesn't mean they have to share a support persona. One spouse's support persona is still Dale and the other spouse's support persona is still Alex (a support persona for the other spouse before the two spouses met each other), but both Dale and Alex have access to all group device histories for both spouses. , but either Dale or Alex will be on the other end of the phone depending on who is calling. Aspects of the present disclosure include implementations where both spouses make a conference call simultaneously on the phone with both Dale and Alex, and the personalities blend naturally. So, IS creates an "enhanced" customer support experience for you and your wife.

According to aspects of the present disclosure for data management as shown in FIG. 12 , there are elements that create a reference persona on the fly or in a repository of reference persona 1200 , an IS instance from one reference persona in each new context. 1201 is generated. This instance can be saved and updated after each interaction, or it can be dynamically recreated whenever needed based on parameters from previous interactions. In some embodiments, IS instances are cached for a limited period to eliminate latency, but parameters are stored, so that they can be reconstructed exactly where they left off, even when offline for an extended period of time.

Training of intelligent systems

Machine learning (deep neural networks, machine learning, CNNs, RNNs, GANs, etc.) may be implemented. It can be implemented to capture and classify these behavior-specific biases and then imitate them in "human behavior". Referring to FIG. 13 , some of the other layers of basic personality 1300 have been discussed above evolution filters, relationship filters, and behavioral masks. The next layer in mapping human cognitive behavior is to map behavior-specific biases. To train an IS to behave as a human, there are a number of steps that can be taken. Behavior collection 1301 begins with an expert system built by psychologists based on our knowledge of cognitive biases 1302 . It is augmented, enhanced and mostly replaced by behavioral data 1303 observable in both the human world and the human/IS virtual world. Observable behavioral data can be generated by observing conversations in the human world. Based on the model of cognitive biases and psychological profile, there are expectations of how a person will respond in a particular conversational setting, and the model is updated when the response differs from the model. In this situation, the dialogue setup can be created by passively observing the dialogue between humans with known psychology, or it can be actively generated through dialogue between humans and the IS. IS can present or discuss topics with humans with known psychological profiles and measure human responses based on predicted responses. The predictive model may be updated based on actual human responses. Next, the behavioral analysis is mapped to cognitive biases 1304 . The resulting behavioral biases are useful in imputing how the IS responds if they are a particular type of individual 1305 (based on all strata above) and how they apply to other behaviors and choices 1306 . used The combination of individual behavioral expectations 1305 and environmental choices 1306 is applied as a function to generate behavioral bias 1307 of the IS. The behavior and task of functions 1308 can be observed and learning is fed back into observable behavioral data 1303 . Once instances of the IS are up and running, they can continue to evolve by starting training each other with GANs (Generative Alternative Neural Networks).

general neural network training

In accordance with aspects of the present disclosure, an IS system may include one or more of several different types of neural networks and may have many different layers. By way of example and not limitation, a classification neural network may consist of one or multiple convolutional neural networks (CNNs), recurrent neural networks (RNNs), and/or dynamic neural networks (DNNs).

Figure 29a shows the basic form of an RNN with a hierarchy of nodes 2920, each of which is an activation function (S), one input weight (U), an iterative hidden node transition weight (W), and an output transition weight (W). (V) is characterized. The activation function (S) may be any non-linear function known in the art and is not limited to a function (hyperbolic tangent (tanh)). For example, the activation function S may be an S-phase (Sigmoid) or ReLu function. Unlike other types of neural networks, RNNs have one set of activation functions and weights for the entire layer. As shown in FIG. 29B , the RNN can be considered as a series of nodes 2920 with the same activation function traveling through times T and T+1. Thus, the RNN maintains historical information by feeding the results from the previous time T to the current time T+1.

In some embodiments, a convolutional RNN may be used. Another type of RNN that may be used is a Long Short-Term Memory (LSTM) neural network that adds a block of memory to an RNN node having an input gate activation function, an output gate activation function and a forget gate activation function, which are described herein. It allows the network to retain some information for longer periods of time, as described by “long-term memory” neural computation 9(8):1735-1780 (1997) by Hokrater and Schmidthuber, incorporated by reference.

29C shows an example layout of a convolutional neural network, such as a CRNN, in accordance with aspects of the present disclosure. In this figure, a convolutional neural network is created for an input 2932 having a size of 4 units high and 4 units wide, giving a total area of 16 units. The illustrated convolutional neural network has a filter 2933 of size 2 units high and 2 units wide with a skip value of 1 and a channel 2936 of size 9. For clarity, only the connections 2934 between the first column of channels and their filter windows are shown in FIG. 2C . However, aspects of the present disclosure are not limited to these implementations. According to aspects of the present disclosure, a convolutional neural network implementing the taxonomy 2929 may have any number of additional neural network node layers 2931 , including additional convolutional layers of any size, fully connected layers, and pooling layers. may include these layer types such as s, max pooling layers, local versus normalization layers, and the like.

과

사이에 분포된 랜덤 값들을 가져야 하며, 여기서, n은 노드에 대한 입력들의 수이다.As can be seen in FIG. 29D , training the neural network (NN) starts with weight initialization of the NN (2941). In general, the initial weights should be randomly distributed. For example, an NN with a tangent activation function is

and

It should have random values distributed between, where n is the number of inputs to the node.

After initialization, the activation function and optimizer are defined. Then, the NN is provided with a feature vector or an input data set. Each of the different feature vectors can be generated by NN from inputs with known labels. Similarly, NN may be provided with feature vectors corresponding to inputs with known labeling or classification. The NN then predicts a label or taxonomy for the feature or input 2943 . The predicted label or class is compared to a known label or class (also called ground validation) and the loss function measures the total error between the predictions and the ground test for all training samples 2944 . By way of example and not limitation, the loss function may be a cross-entropy loss function, a quadratic cost, a triplet contrast function, an exponential cost, and the like. Several different loss functions may be used depending on the purpose. By way of example and not limitation, for training classifiers, a cross entropy loss function may be used, whereas for pre-trained embedding learning a triplet contrast function may be used. The NN is then optimized and trained (2945) using the result of the loss function and using known training methods for the neural network, such as backpropagation, such as adaptive gradient descent. At each training period, the optimizer tries to choose the model parameters (ie, weights) that minimize the training loss function (ie, total error). The data is split into training, validation and test samples.

During training, the optimizer minimizes the loss function of the training samples. After each training period, the mode is evaluated on the verification sample by calculating the verification loss and accuracy. If there is no significant change, training can be stopped and the resulting trained model can be used to predict labels in the test data.

Thus, a neural network can be trained from inputs with known labels or classifiers to identify and classify these inputs. Similarly, a NN can be trained using the described method to generate feature vectors from inputs with known labels or taxonomy.

generative pairwise NN training

To train a generative allele NN (GAN) layout, two NNs are required as shown in FIG. 30 . The two NNs are a first NN 3002 that generates a synthetic source reaction 3005 from a source reaction 3001 and a target reaction 3005 and classifies the reaction 3006 as a target reaction 3004 or not. The second NNs are set opposite to each other. The first NN 3002 is trained based on the taxonomy made by the second NN 3006 (3008). A second NN 3006 is trained based on whether the taxonomy has correctly identified the target response 3004 (3009). A first NN 3002 , hereafter referred to as a generative NN or G _NN , takes the input responses z and maps it to the _{expression G(z; θ g ).}

The second NN 3006 is hereinafter referred to as a differential NN or D _NN . D _NN takes the unlabeled mapped synthetic source reactions 3006 and the unlabeled reaction (x) set 3004 and attempts to classify the reactions as belonging to the target reaction set. The output of D _NN is a single scalar representing the probability that the response will come from the target response set 3004 . D _NN has a data space D(x; θ _d ), where θ _d represents the NN parameters.

The NN pairs used during the training of generative opposing pair NNs may be multilayer perceptrons similar to the convolutional networks described above, but with each layer fully connected. A generative allele NN is not limited to a multi-layer perceptron, and may consist of CNNs, RNNs, or DNNs. Additionally, a pairwise generative NN may have any number of pooling or softmax layers.

During training, the goal of _{G NN} 3002 is to minimize the inverse result _{of D NN .} That is, G _NN is trained to minimize log(1-D(G(z))). _{Since the D NN} is very different from the target response set at the beginning of training, problems may arise in which input responses mapped with high confidence are rejected. As a result, the equation log(1-D(G(z))) saturates quickly and learns slowly. To overcome this, initially G can be trained by providing a much stronger gradient at the beginning of learning and maximizing log D(G(z))) with the same fixed dynamic point. Additionally, GANs are available at https://arxiv.org/pdf/1703.10593.pdf (30 Aug 2018), incorporated herein by reference, in Zhu et al., “Images using period-consistent allele neural networks Unpaired image transformation to" ArXiv, ArXiv: 1703.10593v5 [cs.CV] can be modified to include a periodic coincidence loss function to further improve the mapping results discussed.

The _{purpose of training D NN} 3006 is to maximize the probability of assigning the correct level to the training data set. The training data set includes both mapped source responses and target responses. D _NN provides a scalar value representing the probability that each response in the training data set belongs to the target response set. Thus, during training, the goal is to maximize log G(x).

The first and second NNs together form a minimax game for two with the first NN 3002 that seeks to generate a response to fool the second NN 3006 . The equation for the game is: min _G max _D V(D,G) = E _x~pdata (x) [log D(x)] + E _z~pz (z) [log 1- log D(G() z))]

G _NN and D _NN are trained in a stepwise fashion by _{optimizing D NN and} then optimizing G _NN. This process is repeated several times until no further improvement is seen in the discriminator. This occurs when the probability that the training response is the mapped input response (p _z ) is equal to the probability that the training response is the source response (p _{data ).} That is, when p _z = p _data , alternatively D(x) = ½. Similar to those discussed above for neural networks in general, G _NN and D _NN may be trained using mini-batch stochastic gradient descent or any other known method for training compatible neural networks. For more information on training and constructing allelic generative neural networks, see Goodfellow et al. "Generative allele networks" arXiv:1406.2661 at https://arxiv.org/abs/1406.2661.

expert system

Expert systems generally use forward chaining or reverse chaining. According to aspects of the present disclosure, some embodiments of an expert system may use forward chaining. Additionally, embodiments of the present disclosure may use prospect theory and generate synthetic data sets to support the development and training of expert systems. As can be seen in FIG. 14 , there is initially a series of cognitive biases 1400 . Although this field continues to expand, there are now 47 common areas: priming, hypocrisy, confirmation bias, post-confirmation bias, sharpshooter error, procrastination, bias toward normality, availability heuristic, bystander effect, Dunning Kruger effect, and apo Peniabrand affinity, argument to authority, argument based on ignorance, scarecrow attack fallacy, personal attack fallacy, equitable world fallacy, public goods game, ultimatum game, subjective verification, cult indoctrination, groupthink, supernormal Releasers, Emotional Heuristic, Dunbar's Law, Sold Out, Self-Oriented Bias, Lighting Effects, Third-Party Effects, Catharsis, Misinformation Effect, Adaptation, Extinction Trigger, Social Roping, Transparency Illusion, Learner Lethargy, Embodied Cognition, Anchoring effects, attention, self-mutilation, self-fulfilling predictions, momentary, consistency bias, representativeness heuristics, expectations, illusions of control, and basic attribution errors. It should be noted that the above list is provided without limitation as the field of known biases is currently expanding and any number of biases may be added with sufficient programming known by those skilled in the art. These biases are identified by experts 1401 . Then (still using experts) a reference value is attributed to each action 1402 . These reference values are then stored (1403) in the data store of the expert system quantification of the reference bias. Ultimately, it doesn't matter whether or not it's completely accurate, as the system will learn more from real-world behavior over time, but it does need a baseline to start the process and provide a lens through which to see the behavior.

Development of an expert system begins with determining cognitive biases, mapping them to contextual environments, and assigning a reference value to each action 1402 . For example, without limitation: To get some baseline figures for apophenia (the tendency to misperceive connections and meanings between unrelated things), psychologists use coincidence (meeting acquaintances or having people in your class have the same birthday). ) and list the 100 most common coincidences. The psychologist then determines the weight of each event's likelihood of predictable surprise from statistical and experimental data. These statistical and experimental data may be generated through investigations or observations of human subjects. These mappings are basic mappings - based only on general expected responses and not on any individual personality at all. The good thing about expert systems is that they are weighted naturally so that the results are not limited to binary results. For example, without limitation, see meeting someone on vacation at home. Due to social connections, general media, general social behavior, climate and vacation schedules, it may not be uncommon to run into people who share common friends, but it is always surprising. For example, at a large resort in Las Vegas, odds might suggest that it will happen once in every 30 social interactions over spring break, but when asked people suggest that it will happen once in 100. Thus, for this particular case of apophenia, an “excess likelihood value” of 3.33 (100/30) would be given. Some number (possibly dozens) of apophenia examples are collected and a set of apophenia criteria for the most common situations are created in the examples. These are stored as data for all other behavior-specific bias criteria. Then, when the expert system is faced with social situations in which apophenia can be applied to humans, the expert system applies criteria appropriate to the situation. A similar approach can be used to define expert system criteria for each behavioral bias.

Create a baseline personality

In order to calculate the effect on the baseline personality, there must first be a way to tell the status of the baseline personality. To this end, a series of matrices are used, where each matrix is a number of dimensions within a social or psychological domain, and each layer is a mask or function acting on the layer below it. 15 shows a block diagram of the components that make up the personality criterion 1500 . a base layer 1501 consisting of DNA, RNA, gender, physical attributes, Myers Briggs, five factors, and personality traits; Cultural hierarchy (1502) consisting of parenting, country, state, city, region, religion, culture, family structure, etc.; Training hierarchy (1503) consisting of early learning environment, parenting, learning focus, education, work experience, etc., country, town There is a general environment layer 1504 composed of , physical environment, etc., and a specific environment layer 1505 composed of social environment, weather, time and other related factors.

Regarding the robustness of the data in each matrix, the descriptions need not be completely accurate. If someone is described as 13% on the scale of narcissism, it doesn't matter if that's an exact number because that number isn't a point. Since the goal is to obtain personality outcomes, the system will learn over time to quantify personality factors based on actual behavior-specific outcomes and adjust the weights of the various parameters that make up matrices, masks and functions over time. will be. The training set is initially a behavioral analysis described by an expert system. An expert system is linear where each link in the forward chain derives from its predecessors. However, in the psychological system, there are many factors and not necessarily decisive. Instead, hundreds of other inferential factors may be associated with each expected behavior. For example, without limitation, components of narcissism may be; 1) there is an exaggerated sense of self-esteem; 2) there is a sense of need for qualified, sustained, and excessive praise; 3) expect to be recognized as superior without achievements that warrant superiority; 4) exaggerate achievements and talents; 5) is preoccupied with fantasies of success, power, prowess, beauty, or the perfect mate; 6) believe they are superior and can only associate with equally special people; 7) monopolize the conversation and belittle or ignore those deemed inferior; 8) expect special favors and unquestioned compliance with their expectations; 9) use others to get what they want; 10) inability or reluctance to recognize the needs and feelings of others; 11) envy others and believe that others envy them; 12) behave arrogantly or arrogantly, be arrogant, proud, and pretentious; 13) insist on the best of everything. If a person has strong associations with all these components, then they are definitely narcissistic. However, if you have half of the components in moderation, they are still narcissistic, but to a lesser extent. Conversational analysis can be performed for people whose psychological profile is known, in this case 13% narcissistic. General responses and cues will be determined from these conversations by psychologists, initially, using a defined set of cues. Thus, a structured set of principles has been constructed that weights the determination of the degree of narcissism. Similar lists of all other psychological components are constructed. Nearly 13% of all known personalities, who are narcissists, can be observed, determine what other factors make up one's personality, and once other personality traits are discovered, they are mapped back to the scale of narcissism.

The following groups of factors can be used as input vectors for a neural network. A neural network trained using a machine algorithm to predict a label for a set of actions whose label is based on a measure of personality can be used to label groups of factors based on a measure of personality measure. The language that generated these factors (eg, cats in the original training set) is then used to determine which language generates which tendencies with respect to psychological definitions. From there, a language corresponding to different psychological profiles can be generated.

base layer

The first component of the base layer is DNA. The comprehensive simulacrum of humans includes many, if not all, of all factors that affect human personality, including genetic makeup. According to aspects of the present disclosure, the DNA of the base layer can be represented by important known genetic sequences affecting personality or genetically predetermined conditions affecting personality. Information about the DNA of the base layer can be, for example, without limitation, factors affecting personality such as physical gender and gender identity, body type, coordination, visual and auditory sensitivity, and other physical primitives such as predisposition to heart disease or diabetes. can There are also psychophysical primitives such as dyslexia and left-handedness. DNA factors may be the first dimension of the base layer matrix. RNA forms the next dimension of the base layer matrix as these genes may or may not express themselves based on transcription and regulation by RNA. Because RNA expresses itself differently over time - mostly in the early stages of life - it has a dynamic effect. Additionally, some are in early stages of development, such as breastfeeding or sleep training, while early stages have sociological implications. The dimensions of the matrix associated with DNA and RNA can be defined by geneticists and can change with increasing information about how DNA affects personality. The dimensions of the matrix for very early development can be defined by early childhood psychologists. It should be noted that each entry in the matrix must be weighted. Again, you can start with the default weights based on the opinions of experts in the field, but the final weights and their impact on personality will be updated over time based on observations and experience. The next dimension of the base layer matrix is the personality continuum as shown in FIG. 16 . As discussed above, in some implementations in accordance with aspects of the present disclosure, extroverted->introverted 1601, sensing->intuitive 1602, thinking->feeling 1603, judging- > The Myers Briggs Type Indicator (MBTI) expressed as a percentage in the weight matrix 1600 along the axes of the recognition type 1604 provides a set of numbers for the dimensions of the base layer matrix. Another dimension may be attributed to the five personality traits: openness to experience, conscientiousness, extraversion, amicableness, and neuroticism. There is one more dimension, represented by other personality continuums such as Machiavellianism, achievement needs, cognitive needs, authoritarianism, narcissism, self-esteem, optimism, and empathy. All these factors can be used together to create a base layer matrix. This is the expression of personality at the most basic level. The next layer is the cultural layer.

cultural hierarchy

The cultural hierarchy captures how IS is nurtured. This layer contains background information for IS as shown in Fig. 17 and without limitation; In which countries, states, cities and regions (1701) did IS grow? Did IS grow up in the city or in the countryside (1702)? What religion is IS affiliated with (1703)? What was the political climate in which IS grew up (1706)? In what culture did IS grow up (1704) and what was the family structure that was part of its upbringing (1705)?

It is very diverse across cultures. “Asian cultures are more collectivist, and people in these cultures tend to be less extroverted. People from Latin American cultures tended to score higher on openness to experience, while Europeans scored higher on neuroticism.” All of these differences are captured by the cultural hierarchy and contribute to the different characteristics of the IS.

As another example of the impact of culture on personality, people living in collectivist cultures (typically in Asia, Africa, and South America) value social harmony, respect, and group needs rather than individual needs, whereas individualistic cultures (generally People living in North America, Australia, and Europe) display more individualistic personality traits.

There also seems to be regional personality differences within the United States. Researchers analyzed the responses of more than 1.5 million individuals in the United States and found three distinct regional personality clusters: “kind and traditional… ; More relaxed, emotionally stable, calmer, more creative… and stressed, irritable, and depressed”

This cultural data fits into a layer cake of matrices. According to some embodiments of the present disclosure, not all cities and towns, and not all families, need be represented. We only need a sufficient number of examples to generate our personality. For example and without limitation, creating a customer managed IS instance requires only a few representative backgrounds, for example, 100 US cities with a population between 200,000 and 1,000,000 - a representative sample of all US cities of that size. is not needed Looking at the cultural nurture layer cake, the explanation can start with a locale going from less granular to more granular, for example, without limitation, a locale: Continent → Country → County → Town → Region. Psychology and sociology students can map behavioral tendencies to each of these hierarchies. So, for example, based on the collected MBTI data, Nicaraguans are 15% more extroverted than the average and Chileans are 11% more introverted than the global average. The next layer is the city <-> province. There are known differences in attitudes between the city-suburban-rural and the spaces in between. They are recoded into an expert system to begin coupled with the analysis of real conversations (text, phone, movies, etc. - see below), and over time, they capture feedback from users and relationships with others like mind. It is fine-tuned based on

Using a similar approach, various other cultural dimensions can be added: religion, cultural upbringing, family structure and political environment. These are again seeded by sociologists and psychologists, but will be known by analysis of movies and TV shows, chat sessions, text messages, voice messages and emails.

training layer

In this context, training refers to the teaching or learning of IS, which begins at a very early age. This stratum will answer, without limitation, questions about IS such as: From breastfeeding and eye contact and reading to parenting, what is the impact of the early learning environment? As you get older, what's your learning environment like? For example, is it coed? Is the class large or small? Divisive or focused? What about higher education and work experience? How was the IS instance trained? Did she/he go to college? What are their majors? Did you join the boys fraternity/girls fraternity? What were their grades? How about a graduate degree or certificate? What is your previous job history? All of these factors affect the nature of the IS instance. Essentially, a virtual curriculum for an IS instance is created. The training layer doesn't need to generate the actual events for the IS to memorize (e.g., "Remember the great AEPi Halloween beer party of 1995. Jeffrey was too drunk"). It is sufficient to create a rich bible to ensure the same unique character In some implementations, the character bible is created by users or software programmers using a tool that allows the creator to write unique histories for the IS. In other embodiments, ISs make instances of themselves with their bibles based on circumstances.In still other embodiments, salient characteristics can be outlined and the options the IS can choose - For example and without limitation, a hybrid model can be built that will provide for the IS instance the branches in the bible can be chosen but the basic archetype is immutable for the IS.

Generic Environment Layer

The generic environment layer describes the context for the current state of the IS. That is, it describes the events that caused the IS to contact the user. This layer can answer questions such as without limitation: Where does this IS work? Some of the same factors that influenced the early learning stages of life affect the work or play environment. Is your IS instance in a call center, bar, cube, sales representative or law office working as a litigator? In some embodiments, the general environmental layer includes: agriculture, food and natural resources; architecture and construction, the arts, audio/video technology and communications; management and management; education and training; Jae Won; government and public administration; health sciences; hospitals and tourism; human services; information technology; law, public safety, corrections and security; Produce; marketing, sales and service; science, technology, engineering and mathematics; You can start with the basic career taxonomy of transportation, distribution and logistics. Then, overlays on the same cultural and contextual variables used in the parenting context, such as: Continent → Country → County → Town → Region; City → Suburbs → Provincial. Additional quarters can then be added, such as Large ↔ Small Business, Large Building ↔ Small Building; how diverse the labor force (in multiple axes); What is the politics/religion of the company/executives; Etc. The domains and subdomains of a taxonomy can be defined by social scientists, but the categories, their weight and importance all depend on the results of contact with users and how relevant these inputs are to the value of their outputs.

specific environment layer

The next layer is the specific environment layer. Here the latest elements for the creation of IS instance behaviors are captured. This layer can answer questions like the following without limitation: What's the weather like? How is the traffic safety? how is my morning For example, based on ISI's constituted family, they can have a generally happy marriage with two kids who have to go to school, and they can be in the normal distribution of surprise events (child becomes ill, loses homework, etc.) In addition, the mood of ISI and its actions are affected by these preparation factors.

Analysis by Behavior

The next task of training is for the IS to learn human behavior. There are many interactive voice response (IVR) systems in the marketplace today that use automatic speech recognition (ASR) technologies and natural language processing (NLP) to analyze human behavior. The approach here is to use it as a conversation basis. According to aspects of the present disclosure, the IS is configured to have a personal and social context—knowledge about human social behavior—not an immediate context (this person is shopping or at the beach, etc.).

18 shows an overview of collecting and weighting behavior 1800 in accordance with aspects of the present disclosure. Initially the expert system may be prepared with the cognitive biases collected from the expert system 1801 and a corpus of observable behavioral data 1802 may be added. Embodiments of the present disclosure may use a corpora of chat bot data 1803 and analyze a more normal/social conversational environment. These are without limitation; email and text 1804 , social media 1805 , voice mail 1806 , movies, TV shows, and streaming video sources 1807 . Video media is very socially rich and there is a huge amount of data that can be accessed. Since many movies and TV shows are at risk of being out of the ordinary or "normal", we use Human AI Transformers 1808 (for example and without limitation Mechanical Turk or various other networks of Human AI Classifiers). will be used to grade conversations on a normal scale. Later, when the IS chats with humans, it can learn from human reactions. Additionally, you can chat with other IS instances (using GAN), and hire the same person to analyze the conversation. Depending on the particular corpus of communication data, several different approaches will be needed to perform the analyzes. We will then map the behavioral analysis to behavioral biases 1809 .

According to aspects of the present disclosure, movies and TV may be one source of information used to train the IS. They do not necessarily represent typical long-term achievements. Many genres always end happily, others are mean, situational comedies often rely on people lying, and then the comedy follows. However, in many of these same titles, action per minute is very human. Punch lines or shocking events are often surprising and unusual, but everything in between actions is normal. In some embodiments, mechanical turk/classifiers are supplied with tens of thousands of video assets to be able to scale normally. This approach works for movies and TV, but there are other video assets - especially streaming social media videos (eg YouTube, Facebook videos, etc.). In some embodiments, neural network 1808 is used to perform social analysis and filter out “abnormal” behavior. The value of this corpus differs from that of other social analyses (actually each analysis group) by almost definition, and therefore must be kept separately (labeled with its own corpus).

The next data corpus is chatbot data. This is especially relevant when people ask questions and ask for specific answers. Responses from the various chatbot corpus will focus on accuracy. According to aspects of the present disclosure, the IS may be trained to provide the most human answer, not necessarily the correct answer. One important result from chatbot data is when the bot gets it wrong. Since the people asking the bot are real humans, they won't always be satisfied with the "correct" answer (ignoring the moment when the bot misinterprets or misinterprets the question). Thus, in some embodiments, when a customer suffers because the software is not designed to solve the problem they are trying to solve, the responses will often be rather dissatisfying to humans rather than saying "I'm sorry, the software doesn't do that." . A more sensitive bot could say something like, "Let me see if I understand the problem I'm trying to solve." Then, after properly verifying that the bot "understands" the question, you can guide them to the appropriate software to solve the problem (Santa Cruz of Macy's at Miracle on 34th Street who sent the family to a competitor to the brand of Macy's) It's a bit like a big increase in value). Perhaps the bot is "their bot". That is, they stay with the customer to help them install the other software correctly (actually the same bot persona will work for the competition, and relevant pieces of customer data could be shared with permission).

Next could be social media analysis of the list of corpus for behavior analysis. This group has acres of communication data, which are usually curated. People often try to show the best side to their acquaintances and the more sincere side to their close friends. Anonymous or pseudo-anonymous posts can be particularly atypical and should probably not be included at all. Private posts will be particularly valuable because of the natural candor associated with them in general (unnecessary posts like "Another Perfect Day" or "Life sucks as usual" are filtered out). Again, it is bucketed into several separate corpuses and the weight of humanity is defined by the analysis of actual responses to actions.

The last buckets in the corpus are text, voicemail, and email. The text is most natural in terms of tone. In some embodiments, emoji and acronyms extend to actual verbal descriptions of the emotion or context expressed in the emoji or acronym. Knowing the various chat members and the relationships between them will provide context or metadata for each group. People will talk to their mother differently than their friends and they will talk to a group of 4 friends differently. The application of this context will be very important for AI training. Voice mail is, in some ways, a subset of text data. It usually targets a single person or household. Users, or voicemail, are typically a different demographic from that of text messages and may exhibit a different or previously established personality. Young people today rarely use voicemail. Emails may need metadata to describe the situation and subdivide it into multiple corpuses. In some embodiments, a thread may be parsed into another corpus (eg, without limitation, personal communications within business email).

The next task is to correlate the above behavioral observations with a list of behavioral trends (behavioral biases) 1809 . This parsing answers the question of which behaviors are associated with which tendencies. Here experts are used in two stages. Classify the first expert actions into groups. For example, without limitation, there are people who are quick-tempered, empathetic, anxious, or relaxed. Experts create taxonomy of emotional groupings and then human classifiers that label all behaviors. These behaviors and their levels are then used to train machine learning systems (eg, neural networks).

Shifting behavioral biases to individuals

Now everyone has a different degree of behavioral bias. One can be very "analytic" and "realistic" while the other tends to be more emotional and exaggerated. Although prone to exaggeration, it is usually limited to a few domains. A person exaggerating his role on the high school basketball team can be very accurate when he remembers the details of a trip to Greece. Remember that the contextual criterion is essentially the personality of the IS developed in the stack (DNA → early development + gene regulation → basic personality factors → developmental filter → relational filter).

As shown in FIG. 19 , contextual criteria 1900 are mapped against each behavioral bias 1901 to create a contextual set of biases for each IS instance 1902 . By way of example and not limitation, start with one IS instance and then extrapolate to more general instances. For example, an IS instance can be created with the following parameters: A 35-year-old heterosexual woman, raised in a small town in the Midwest. Although she grew up as a Protestant, she did not go to church regularly - however, she has strong fairness values. She has two living siblings, a brother and a sister. Her brother is gay, and this shows her openness to her relationship (not very judgmental). However, she is very critical of abortion and the right to life issues and believes that life should not be wasted as there are enough people who want children. She is an EFTJ (64, 81, 59, 50, Myers Briggs) and the five factors were openness (66%), conscientiousness (34%), extraversion (72%), friendly (48%) and neurotic (32%). , her other personality traits were neuroticism (53%), Machiavellianism (14%), need for achievement (21%), need for cognition (73%), authoritarianism (26%), narcissism (14%), and self-esteem (58%). , optimism (87%) and empathy (13%). She enjoys crossword puzzles and medical dramas. There are potentially many more components in her character bible, but this gives some idea of who IS is as a fictional character. Now, as shown in FIG. 20 , it is possible to map all psychological parameters to a plurality of behavioral biases 2000 , and provide weights to each 2001 . These are reference biases and have not yet been mapped to a specific situation. However, they each have a baseline value, namely, "How distinct is the behavioral bias?" Using several psychologists, we will map Myers Briggs and five factor tendencies to our behavioral biases, as discussed above. This can be done at any scale, but by way of example and not limitation, percentages may be used where higher numbers are more prone to bias and lower numbers are less prone to bias.

As the IS interacts with other users and ISIs, it will become more fully developed and the number, name and weight of the biases will change. Even starting with an initial set of biases, ISI has a basic set of personality elements for how to respond to new situations. We can now begin to apply IS with biases for certain situations. Remember that behavioral bias is only one component of personality, so you need a complete personality bible when preparing to deal with specific situations. For example, the above heterosexual woman's sensitivity to abortion and adoption will greatly affect the way she behaves in certain situations and may not be the right character for a planned parental early pregnancy support group, but a church-based early pregnancy support group. It may be the correct profile for

Transferring behavioral biases to the contextual environment

21 shows the next layer of the stack, mapping the IS instance to the current context. Contextual criteria 2100 and behavioral biases 2101 are mapped to create contextual biases for each IS instance 2102 , which will then impute the biases to contextual environments 2103 .

This tier answers questions such as, without limitation, that IS sells makeup at makeup counters, provides customer support, acts as a health care worker or consultant at a resort, and the like. Some contextual behavioral data is useful here. It can start with an expert system with axes based on psychological data. 22 , at the highest level in region 2200 , it is based on behavior-specific parameters (Siberia is very different from Khan). Everyone is affected to some extent by the time 2201 they are in. Many other parameters are continuous between the two different aspects. For example, without limitation, there are deep social differences between city 2202 and local society 2203 . For this, we can use population density. Also, there is a continuation between Entitled 2204 and Earthy 2205 . People behave one way in Beverly Hills and the other in Compton. How crowded is the environment? Does the IS work in a bustling mall or only from home (this is how you set up personality behaviors - this doesn't actually require any physical instantiation, and in some embodiments the IS will be 'live' in the cloud)? Other factors include, but are not limited to, working or playing in large groups (2206) or solo (2207), what kind of work (2208) they do and what kind of play (2209) they do, work (2210) and what the environment is like 2211 where they play in their spare time, etc. Now, looking at the various environments, we can dig a little bit by type: Insurance office? Graphic design house? Software programmer cube Is it a library or a beach or a cruise ship It will use an expert system created by humans to create a complete taxonomy It doesn't matter whether the system is completely accurate as AI will iterate over time and develop its own taxonomy based on experience not.

Once the basic parameters of the expert system are established from psychological and sociological data, classifiers can map expected behaviors to social environments. In this case, it is more important to collect a large amount of recognized social conventions than to determine the exact expression of social norms. This can be done either by surveying a large population or by voting. After a sufficient number of people (say, about a thousand) have conducted a survey about a particular behavior and environment, the system will have a reasonably accurate social perspective.

By way of example and not limitation, assume that IS is a banker. Now, we don't have any specific interactions with individuals yet, but we know a lot about the environment. Let's say the teller is at a country bank where people tend to be friendly and sociable. Banks aren't usually crowded, so people don't usually have to wait, and transactions usually start with a conversation about the weather or a recent event or fair. Let's say the teller has a fairly long (virtual) commute and is aware of the traffic conditions that day. All of this applies to the baseline behavior of the IS instance in this environment.

data ordering

The next step in this stack is to create the full psychological profile that IS will bring to human interactions. To this end, an array matrix can be created that holds and operates on data, as shown in FIG. 23 . Various data sets are assigned to columns and rows within the array. This can be done in several ways, but let's choose a few arrangements as examples and not by way of limitation. Array 1 is the base layer 2300 of the contextual basis. Factors 2301 are placed in the array to include rows for the Myers Briggs axes, the five factor personality trait axes, and other personality traits followed by the physical attributes through the row, row 1 being DNA, row 2 being RNA, row 3 is the microbiome, row 4 is the gender continuum, and so on.

Now, continue the stack to add the next array. This array is the culture layer 2302 . Starting with the actual location, the rows are: subcontinent, country, state, city and region. Each factor has a percentage 2303 and a weight 2304 . So in the example of religion, ISI can be quite religious - say 73%, but have little effect on its life and thus weight can only be 15%. The following are family components of a cultural hierarchy with a continuum such as intimacy, size, gender composition, parent structure, etc. Now where is Row 4 right for you in the community? How sociable are the people in your community? Are you moving to another community? Do you often eat out? Are the communities tightly connected or loosely connected? It is a cultural hierarchy with rows such as etc. Other cultural factors may be religions with columns representing axes of religious influence, such as orthodoxy, cultural unity.

Cultural hierarchy includes early developmental factors such as family size, number of siblings, and underlying cultural background (religion, cultural group, etc.). This is of course expandable and the rows and columns in this array can vary and expand and collapse.

23 , there is a rough approximation of some fields in the various arrays with the understanding that the Excel snapshot does not convey the array's multidimensionality but still provides some width. Figure 24 illustrates another approach to data visualization and can help clarify the fact that the data consists of multiple arrays.

Using the same line of thinking for the training layer 2305 , a generic environment layer 2306 and a specific environment layer 2307 are created. All these layers will be combined to create IS instances up to the point of human interaction.

An Excel snapshot can tell you some of the width, but it doesn't convey the multidimensionality of the array. 24 depicts another approach to viewing data, which is a more exemplary way of visualizing this and can help clarify the multidimensionality of arrays. 2400 shows magnitude 2401 indicating how strong this personality trait is within the continuum of the personality binary (such as introversion versus extroversion) of this particular IS and weight 2402 shows how much weight this trait weights when making decisions in social situations. indicates how much personality aspect should be taken into account when weighting relevance to a particular situation. Each dimension of the array is associated with a different layer including a base layer 2403 , a culture layer 2404 , a training layer 2405 , a general environment layer 2406 , and a specific environment layer 2407 .

Mapping data to actions

As discussed above, data is captured in chat sessions, text messages, videos, etc. but still needs to be mapped to personality traits to be expressed.

25 illustrates segmentation of a dataset in accordance with aspects of the present disclosure. The collected behavioral data 2500 is initially used by human AI classifiers to monitor and classify behaviors and comments related to movies and TV, chatbot corpus, social media, email and voicemail 2501 . is created when From this corpus, we will create an initial behavioral mapping from human classifiers 2502 .

After the initial set of collected behaviors are analyzed by classifiers, these tagged behavior (conversation, physical behavior, etc.) bits are labeled according to the taxonomy described above. All other elements of the base layer, culture layer, training layer, general environment layer and specific environment layer are labeled. Once the various factors have been analyzed, one or more deep neural networks (DNNs) 2503 can be used to learn these taxonomies.

The DNN may then analyze the second data set 2504 . The accuracy of this mapping can be checked by human AI classifiers (perhaps a subset of the classifiers above marked as more skilled). We will continue this review and iteration until the DNN does as well as the best classifiers (2506). At the same time, a score may be generated for the classifier's ability to classify. This is not necessarily psychological correctness, but popular behavioral correctness. That is, if human classifiers most often agree with the most popular opinion, they score higher as classifiers. Once these few rounds are complete and humans have a good idea of how to classify behavior, predictive accuracy can be contrasted with the well-informed psychological beliefs held by experts in the social sciences and psychology.

Create social taxonomy

From the analysis of conversations labeled by human classifiers, a taxonomy of social situations can be generated. Again, psychologists and sociologists create a standard set of expectations and structure and group them into behaviors (2600, 2601, 2602, 2603) and sub-actions (2604, 2605), and further subdivide them into taxonomies (2606, 2607). can do. And again, these groupings are not completely deterministic, they are just a starting point. The success of any choice is measured by how closely the taxonomy of social situations achieves set goals. This starts to come down to all the biggest questions of how to set, manage, update, and manage goals. This interactive taxonomy will be augmented and continuously adjusted based on advances in natural language processing (NLP), sentiment analysis, and emotion detection. Note that emotion detection is not limited to text and speech, but is making great strides using visual cues.

Generating a full baseline personality

The next step is to take all of this data and use it to define the overall baseline personality. Fig. 27 is built on top of Fig. 24; This is the nature of ISI when humans first interact with it. So far we have created a very wide and deep matrix of factors 2700 . Initially, we will create dashboards that humans can use to assemble personalities. All data for the three initial tiers (ace 2703, culture 2704 and training 2705) and two environments (general 2706 and specific 2707) are faders or other to change values. It can be arranged as a series of columns and rows with input types, which can be tested by people for basic functionality. At this point, IS instances are fairly logical. They will still react with warmth and humor and all sorts of other human emotions, but these will be quite (and perhaps too) logical. After generating the overall criterion personality, we will add cognitive biases or behavioral masks 2708 and evolution filters 2709 and relational filters 2710 .

How to develop a corpus of situations in which to operate our functions (bias)? A very limited subset of situations for imputing biases can be used initially and over time the system can train to impute biases to more and more situations (as humans do).

Limiting the scope of the initial interactions will form the basis for development. Customer support over the phone may be one of the first areas of development. This is due to a number of reasons, including historical data from many corpus, including experiences of communicating with customers. The first goal may be for one ISI solving one customer's problems. The ISI can be prepared as a corpus of support conversation data to be already in the current level of chatbots. In addition to this, the first set of psychological primitives may be overlapped to add humanity to the interaction. Basic human interactions (hello, how are you, how can I help, etc.) may be early psychological primitives used. At the same time, the psychological meaning associated with the questions and conversations can be measured. In the first stage (voice-only customer support), although the ability to identify emotions/intents is limited, only the addition of voice analysis allows the identification of communication through sound and visual or other perceptions of the IS to understand and respond. It will greatly increase your ability. The next step will be video chat. Ultimately, other senses can be added to function similar to physical beings (pets, robots, etc.).

Returning to the full context-aware profile, there are five layers (base - 2703, culture - 2704 and training - 2705), two filters (development - 2709 and relation - 2710), two environments (general - 2706 and specific - 2707) and masks (by behavior - 2708), there is a large set of elements.

As can be seen above, there are many variables. However, grouping and subgrouping algorithms can be used to control the dashboard to set variables for various parameters of the personality (ISI) for these layers, filters and masks, corresponding parameters as shown below. can be used to construct appropriate responses in real time.

Once the ISI is generated, it can be tested by human classifiers capable of judging the ISS, and the judgment can be used to refine the ISI. Once a reasonably good ISI set is established, they can be trained on each other. The human classifiers can then be used again to judge the outcome of unsupervised training, and more insight can be gained by observing the unsupervised training process.

Create an action-specific function

ISI characteristics can be described using a multidimensional matrix. For convenience, the matrices described herein are 16 x 2 x 8 (e.g., extroverted (2800), sensing (2801), thinking (2802), judging (2803), open (2804), conscientious (2805), Friendly (2806), neurotic (2807), Machiavellianism (2808), achievement (2809), cognition (2810), authoritarianism (2811), narcissism (2812), self-esteem (2813), optimism (2814), and empathy (2815) ) x (size 2816 and weight 2817) and 8 dimensions (layers, filters, and masks) psychologists and sociologists (to keep the matrix dimension for matrix math simple) 16 We will select the most important factors: Letters can be used to represent individual dimensions or layers, filters and masks: B (Base Layer - 2818), C (Culture Layer - 2819), T (Training Layer - 2820) ), D (Evolution Filter - 2821), R (Relationship Filter - 2822), G (General Environment Layer - 2823), S (Specific Environment Layer - 2824), and B (Behavioral Mask - 2825). It should be understood that it can be of any size configured to describe the personality traits of the IS now produces a representation of the personality in which the IS can act Convolve the layers in a way not dissimilar to the way convolutional neural networks (CNNs) work Aggregation hierarchies can be created by routing, see the matrix in Fig. 28, "Multidimensional Matrix Mathematics: Notation," by Ashu MG Solo et al. available at http://www.iaeng.org/publication/WCE2010/WCE2010_pp1824-1828.pdf. Representation and Simplification, Part 6 of 1”, an approach to representation of multidimensional matrices, the contents of which are incorporated herein by reference.

Now, the representation of the reference personality has been created. As mentioned above, this is the nature of the IS instance prior to interacting with any human. Some human interactions will create changes to the IS instance and should be saved as a new instance or as described modifications to an existing instance. There is a set of behavior-specific biases indicated above and a set of situations in which these behaviors occur. A psychological description of the IS instance is now established, and behavioral biases can be applied to each interaction. The defined situations can be mapped to the relevant biases through the lens of the IS instance. The three-dimensional personality matrix can be expressed as Ф(Phi). Each of the 48 behavioral (cognitive) biases has two factors: magnitude and weight (48 x 2 matrix). The behavior-specific bias matrix can be expressed as Λ (lambda), and the additive matrix calculation can be performed in each situation as follows: f (behavior) = {Φ}·{Λ}.

feedback loop

The feedback loop is very important. ISI can make billions of small decisions, and it is important to learn how well each of these decisions did. To this end, the system can monitor behavioral cues and use them as a measure of ISI performance. Some examples of overt indicators of unsuccessful success include, without limitation: delayed response (not including overly long delays indicative of absence), calling back on the same topic (using text analysis to see if something is not understood OK), anger, ignore, etc.

At a deeper level, a system in accordance with some aspects of the present disclosure may monitor emotional sentiment (along with visual and voice analysis). In particular, the system may be looking for empathy, serenity and engagement. Using these metrics, ISI sets goals in advance to determine how successful it is at the task. Perhaps the best goals are empathy, lack of tension, participation, and (true) words of appreciation, but any set of goals can be chosen depending on the desired outcome (for example, perhaps ISI creators are asking if someone helps the government for hurricane victims). I don't want to get angry about not doing it). If certain responses or approaches are found leading to undesirable outcomes, those approaches or responses may be used less or obsolete, and if certain approaches lead to success in reaching our goals, ISI will be much more More can be altered to behave in such a way. Further, in some embodiments, certain personalities do not work well with certain human personalities to adjust the personality of the IS instance or entirely new personalities - especially when the human personality is known or imputed based on behavior - may be attempted. it can be seen that

Bible Software and Dashboard

In some embodiments of the present disclosure, a dashboard may be created from various personality parameters to allow a human to try out various variables of ISI personalities. In other embodiments, the IS may select its personality parameters based on contextually relevant entities. Of course, personality designers can manually set more than 256 variables and then take an IS instance for a test conversation or interaction. In some embodiments, the IS instance may be modeled on a known entity such as, for example, without limitation, Abraham Lincoln or Catherine Hepburn or Meredith Gray of Gray's Anatomy. In still other embodiments, users may request a specific personality for the ISI or in other embodiments known personalities may be mixed, for example, without limitation; Winston Churchill mixed the voices of Diane Sawyer and James Earl Jones with the mannerisms of Harry Potter. In this way, unique, fun and interesting virtual personalities can be created through real human-like interactions.

reactions

Now that the various filters, masks and functions have prepared the appropriate reaction, our IS must respond in real time in a reliable way. In some embodiments, programmed delays may be added to mask the computation time for the response. For example, after being asked a question, the ISI may respond with an immediate "yes" or "yes" while calculating a longer response. Further, in accordance with some aspects of the present disclosure, bodily responses: fine facial expressions and other body movements, voice tone, breathing, sweat, skin color (blood flow), etc. are appropriate responses that enable the IS to understand human emotions. is mapped to

applications

According to aspects of this disclosure, ISI may be implemented in virtual environments such as video games and textual help lines. In addition, other new virtual environments allowing more 'real' interactions with ISI could be created where the ISI could function in more traditional human roles, for example and without limitation, the ISI could be a stock trader, custodian, could be a doctor ISI can interact with the wider world through virtual environments. Additionally, in some embodiments, this technique may be used with or without VR glasses or room for training.

According to some aspects of the present disclosure, as more and more virtual characters are created in the online and in-game worlds, they will interact and become part of a virtual social network. Real users can engage in social networks - sharing stories, photos, videos, and more. However, the virtual characters ISI may join their own social network but may also be part of a social network populated with a combination of real or virtual characters.

internationalization

The systems described above and below will function in any cultural setting. However, results will vary according to cultural circumstances. Some cultures (like the Netherlands) can be very candid and rude, while others (like Japan) are very subtle and context-dependent. Psychologists, sociologists, and others in the art should isolate the results for each cultural subgrouping and the IS will have to function within that cultural environment.

system

Fig. 31 shows figures throughout the specification, for example an intelligent agent system for implementing methods such as those shown in Fig. 5, Fig. 10 or Fig. 13 for example. The system can include a computing device 3100 coupled to a user input device 3102 . User input device 3102 may be a controller, touch screen, microphone, keyboard, mouse, joystick, or other device that allows a user to input information, including sound data, into the system. The user input device may be coupled to the haptic feedback device 3121 . The haptic feedback device 3121 may be, for example, a vibration motor, a force feedback system, an ultrasonic feedback system, or a barometric feedback system.

Computing device 3100 may include one or more processor units 3103, according to well-known architectures such as, for example, single-core, dual-core, quad-core, multi-core, processor-coprocessor, cell processor, and the like. can be configured. The computing device may include one or more memory units 3104 (eg, random access memory (RAM), dynamic random access memory (DRAM), read-only memory (ROM), etc.).

The processor unit 3103 may execute one or more programs, some of which may be stored in the memory 3104 , the processor 3103 being operable to the memory, for example by accessing the memory via the data bus 3105 . can be tightly coupled. A machine learning algorithm 3121 configured to label and weight the behavioral data and behavioral biases collected in the database 3122 and refine the reference personalities 3109 and IS instances 3108 as discussed above. may include. Additionally, memory 3104 may be configured to generate a response from personality biases and behavior-specific biases stored in database 3122 or as part of reference personalities 3109 , which may be configured to generate one or more expert systems 3110 . can have These responses may also be part of the IS instance 3108 . Database 3122 Criteria Personalities 3109 IS Instances 3108 and Machine Learning Algorithm 3121 Data in Mass Storage 3118 or on a server coupled to Network 3120 Accessed via Network Interface 3114 3118 or as a program 3117 .

Input video, audio, tactile feedback, smell, taste and/or text may be stored as data 3118 in mass storage 3115 . The processor unit 3103 is further configured to execute one or more programs 3117 stored in the mass storage 3115 or in the memory 3104 to cause the processor to perform one or more of the methods described above.

Computing device 3100 is also known as input/output (I/O) 3107 , circuits, power supplies (P/S) 3111 , clock (CLK) 3112 , and cache 3113 . support circuits, which can communicate with other components of the system via, for example, bus 3105. The computing device may include a network interface 3114 . The processor unit 3103 and the network interface 3114 may be configured to implement an appropriate network protocol for the PAN, for example, a local area network (LAN) or a personal network (PAN) via Bluetooth. The computing device may optionally include a mass storage device 3115 such as a disk drive, CD-ROM drive, tape drive, flash memory, etc., the mass storage device may store programs and/or data. The computing device may also include a user interface 3116 to facilitate interaction between the system and a user. The user interface may include a monitor, television screen, speakers, headphones, or other devices that communicate information to the user.

The computing device 3100 may include a network interface 3114 that facilitates communication over an electronic communication network 3120 . The network interface 3114 may be configured to implement wired or wireless communication over a local area network and a wide area network such as the Internet, such as the Internet. The device 3100 may send and receive data and/or requests for files via one or more message packets over the network 3120 . Message packets transmitted over the network 3120 may be temporarily stored in a buffer of the memory 3104 . The classified behavioral database may be available via network 3120 and partially stored in memory 3104 for use.

While the foregoing is a complete description of a preferred embodiment of the present invention, it is possible to use various alternatives, modifications and equivalents. Accordingly, the scope of the present invention should not be determined with reference to the foregoing description, but should instead be determined with reference to the appended claims, along with the full scope of equivalents. Any feature described herein may be combined with any other feature described herein, whether preferred or not. In the following claims, the indefinite article "a" or "an" refers to the quantity of one or more of the following, except where expressly stated otherwise. The appended claims are not to be construed as including means + limitations of functions, unless expressly recited in a given claim using the phrase "means for".

Claims (20)

A method for training an intelligent agent, comprising:
a) generating a personality matrix;
b) combining a cognitive bias matrix with said personality matrix;
c) generating a behavioral function for the situation based on the combined cognitive bias matrix and personality matrix. The method of claim 1 , wherein the personality matrix comprises parameters corresponding to at least a Myers Briggs type indicator. The method of claim 1 , wherein the personality matrix comprises parameters corresponding to five major personality traits. The method of claim 1 , wherein the personality matrix comprises at least one biographical parameters. The method of claim 8 , wherein the one or more electrical parameters include a location, history, cultural or educational parameter. The method of claim 1 , wherein the behavior-specific function is adapted from classified human interaction data, wherein the human interaction data is classified based on parameters of the personality matrix. The method of claim 10 , wherein the behavior-specific function is generated by a neural network trained on the classified human interaction data, and wherein the neural network training is modified by the personality matrix. A method for using an intelligent agent, comprising:
a) determining the situation;
b) applying a behavioral function to the situation based on a combined cognitive bias matrix and a personality matrix to generate a response to the situation. 9. The method of claim 8, wherein determining the situation comprises using at least a neural network trained with natural language processing data. The method of claim 9 , wherein the natural language processing data includes at least one dictionary answer response to a user question. 10. The method of claim 9, wherein the behavior-specific function modifies a response determined by the neural network. The method of claim 8 , wherein the personality matrix and the cognitive bias matrix include at least a magnitude and a weight. 9. The method of claim 8, further comprising c) monitoring for signs of success of the reaction. The method of claim 13 , wherein monitoring for signs of success comprises determining emotional sentiment using machine vision or voice analysis. 14. The method of claim 13, wherein monitoring for signs of success comprises monitoring a delay in a user response, a call back to the same subject, ignoring the user, or monitoring the user's anger. 14. The method of claim 13, further comprising adjusting the cognitive bias matrices and the personality matrices using the indication of success of the response. 9. The method of claim 8, wherein determining the situation comprises using visual information from machine vision or object recognition. The method of claim 8 , wherein the behavior-specific function is applied by a neural network trained on classified human interaction data and the neural network training is modified by the personality matrix. In the intelligent agent system,
processor;
a memory coupled to the processor;
Upon execution, the processor:
a) determining the situation;
b) non-transitory instructions embodied in the memory to perform methods comprising applying a behavior-specific function to the situation based on the combined cognitive bias matrix and personality matrix to generate a response to the situation. Including, intelligent agent system. Non-transitory instructions embodied in a computer-readable medium, which, when executed, cause a computer to:
a) determining the situation;
b) applying a behavior-specific function to the situation based on the combined cognitive bias matrix and personality matrix to generate a response to the situation.

Images